Performance measures from prediction- realization tables

A standard output from the application of probit estimation is a prediction–realization table, which compares the predictions from the estimated equation to the realizations. The percentage of correct predictions, although commonly reported, can be a misleading guide to prediction performance. This paper considers a number of other methods for measuring performance, all calculated in Pseudo-R2 form and all based only on prediction table information. Monte Carlo results are provided to permit comparisons. It is argued that one criterion to choose a Pseudo-R2 is its closeness to the (unobserved) R2 from the underlying latent variable model. A measure suggested by McFadden, Puig and Kirschner scores well under this criterion.